"Big model + small model", the "pragmatic" innovation path of China Building Materials Information

In the past two years, from ChatGPT to DeepSeek , technologies represented by general large models have flourished, bringing unprecedented prosperity to AI applications. However, large models are not omnipotent. Limited by their ability to understand scenes in depth in the industry, the real application potential of large models in vertical fields has not yet been fully explored.

Wang Qiaochen, deputy general manager of CNBM Information Technology Co., Ltd., recently said, " The implementation of AI in industrial scenarios is not about pursuing the most advanced algorithms, but about finding the most reliable solutions."

Indeed, big models should be a component of AI solutions rather than the solution itself. For example, in professional scenarios such as production safety, using mature CV small model reasoning to achieve all-weather safety monitoring, and using big model reasoning to cross-check the detection results, is obviously more accurate and more cost-effective than a simple big model solution.

At the beginning of this year, DeepSeek emerged and quickly triggered the pursuit of AI applications in all walks of life.

Wang Qiaochen pointed out, " The open source of DeepSeek R1 has set off a profound change in the field of AI . It has transformed large models from high-end resources of a few companies into tools available to the public." The most direct impact of this technological equality is that it lowers the threshold for companies to apply AI , allowing industrial scenarios that were previously limited by computing power requirements and closed-source ecosystems to try large model technology more conveniently.

In fact, the essential proposition to be solved when big models are implemented in industry scenarios is the implementation of the "last mile" between technological advancement and industry universality. This seemingly simple logic actually hides complex industry know-how issues.

For example, when focusing on industrial scenarios, especially the key area of ​​production safety, challenges arise. Industrial production has extremely high requirements for stability and reliability, and does not tolerate any mistakes. At the same time, the " illusion " problem of large models may lead to wrong decisions and seriously threaten production safety. In addition, industrial data is scattered and has different standards. How to effectively integrate and utilize this data so that the model can better understand and handle problems in production scenarios is also an important challenge for the implementation of AI in industrial scenarios.

As Wang Qiaochen said, "There is still a huge gap between the amazing performance in the laboratory and the actual needs of the industrial site." For example, in an industrial process such as cement production, a tiny error may lead to a serious safety accident, which requires the AI ​​system to achieve near-perfect reliability. " 99% accuracy is far from enough, we need more than 9 after the decimal point ."

Therefore, to fully utilize the advantages of large models in industrial scenarios is not entirely a technical issue of large models, but a comprehensive issue of combining models with knowledge, experience, model algorithms and scenario understanding. Facing this challenge, the CNBM Information Digitalization Team has explored a unique solution: system engineering capabilities.

Qian Xuesen's "System Theory" believes that a system is a whole composed of a number of interrelated, interactive and mutually influential organizational parts and has certain functions. From the perspective of the system, studying any part alone cannot answer the question of the system's integrity.

This means that the key to promoting the implementation of AI in industrial scenarios is not a single technology issue, but a systematic problem covering many influencing factors. The solution proposed by Wang Qiaochen, "system engineering capabilities", focuses on three aspects:

First, the knowledge base is constructed.

Large models sometimes experience hallucinations in industrial scenarios, but rich and high-quality industry development data can effectively reduce the occurrence of this problem. Through a large amount of mechanism knowledge and production data in the building materials industry, a professional "knowledge base" can be provided for general large models to effectively reduce hallucination problems. When the knowledge base is rich enough and of high quality, the large model can process information more accurately and reliably, thereby improving its application effect in the field of production safety.

Secondly, it is a pragmatic human-machine collaboration model.

Although large models have powerful capabilities, they also have their limitations, and human experience and wisdom are particularly important in industry scenarios. Through human-machine collaboration, human experience and judgment are combined with the powerful computing power of large models: humans can use their experience and intuition in the field of production safety to supplement and guide large models, while large models can quickly process large amounts of data and provide a basis for human decision-making. "For example, simple technologies such as electronic fences, combined with appropriate human-machine collaboration models, are often more practical than complex algorithms." Wang Qiaochen said.

Finally, it is the iterative strategy of taking small steps and running fast.

Under the rapidly changing market demands, timely response and adjustment of R&D paths are key. By taking small steps and running fast, we can quickly verify the value of products and reduce R&D costs and risks. Wang Qiaochen said frankly, "My requirement for the team is that all R&D cycles should not exceed one month. In this way, even if the technology is updated or the big model "crushes" us, the value of our investment has been realized."

It is based on this set of systematic engineering capabilities that the CNBM Information Digitalization Team has proposed a "big model + small model" solution for the field of production safety. In this system, the small model is responsible for high-certainty basic detection tasks, such as helmet wearing recognition; the large model handles complex situations that require reasoning capabilities, such as abnormal judgment in dynamic environments. The two work together to build an efficient and reliable production safety AI system, providing strong protection for production safety.

In fact, the implementation of AI solutions in the field of production safety has always been challenging. Industrial production has a very low tolerance for hallucinations of large models. Although the small models in traditional CV solutions can solve the problem to a certain extent, their accuracy and generalization capabilities are limited.

Therefore, the solution of large model + small model is a more effective AI implementation strategy.

Wang Qiaochen believes that "sometimes, training a large industry model may not be the best choice. As the capabilities of general large models improve, specially trained industry models may soon be surpassed." Therefore, the strategy of China National Building Materials Information is to improve the applicability of scenarios through industry knowledge base and system engineering based on general models.

The advantages of this approach are obvious: it avoids high training costs and can quickly respond to technology iterations. "The general large model is updated almost every week, and we must remain flexible, " Wang Qiaochen explained.

It is not difficult to conclude that CNBM's success in the industrial field is largely due to its focus on the building materials industry, turning industry know-how and data accumulation into a moat. Through long-term industry focus, CNBM has accumulated rich industry data and expertise, which have become the cornerstone for developing precise AI solutions.

At the same time, during the service process, CNBM Information adopts a companion service model. A professional team composed of industry experts, algorithm engineers and operation and maintenance teams tracks customer needs throughout the process and performs multiple model iterations and optimizations every month based on actual production conditions to ensure that the AI ​​system is deeply integrated with the production process, providing customers with continuous and efficient services.

In fact, this service model forms a positive cycle by continuously exploring customer needs and enriching the industry knowledge base. The model algorithm can be continuously optimized, and the system will evolve to be more intelligent.

In general, system engineering capabilities are the core support for the implementation of AI technology in vertical fields such as production safety. The concept of " big model + small model " provides a pragmatic innovation path for industrial AI . The deep integration of AI technology and system engineering thinking is not only a unique competitive advantage of China Building Materials Information, but also provides new ideas for the implementation of AI in vertical fields.

The true value of AI technology lies not in its advancement, but in whether it can solve practical problems. The practice of CNBM Information shows that by building system engineering capabilities and combining the fusion strategy of "big model + small model", AI can be efficiently implemented in vertical fields such as production safety .

It also indicates that a pragmatic innovation model guided by scenario needs and driven by system engineering thinking is expected to become a reference path for the industry's intelligent transformation.